Eval Run

Purpose / When to Activate

Activate when:

• The Discovery Agent runs the Skill Optimize protocol and needs to score a skill output
• A skill's instructions have been modified and a before/after quality comparison is needed
• A baseline score is being established for a skill that has never been evaluated

This skill is generic: it accepts any output file and any evals.yaml, regardless of skill domain.

It follows the GAAI principle "skills never chain" — it evaluates the output it receives; it does not invoke the skill that produced the output.

---

Process

Step 1 — Load inputs

1. Read the output_file path. Confirm the file exists and is non-empty. If missing: FAIL immediately with error "output_file not found: {path}".
2. Read the evals_file path. Confirm the file exists and is valid YAML. If missing: FAIL immediately with error "evals_file not found: {path}".
3. Parse the evals.yaml structure. Validate:
   • skill, version, description, and assertions fields are present
   • assertions list is non-empty
   • Each assertion has id, type, and description fields
   • If any required field is missing: FAIL with error "evals.yaml validation error: {details}"

For the full evals.yaml format spec, see references/evals-format.md.

Step 2 — Run code assertions

For each assertion where type: code:

1. Read the check field. Execute the corresponding mechanical verification:

   check	Verification method
   word_count	Count whitespace-separated tokens in the output file. Compare against params.min and params.max.
   char_count	Count all characters in the output file. Compare against params.min and params.max.
   regex_match	Apply params.pattern as a regex to the full output text. PASS if at least one match found.
   regex_not_match	Apply params.pattern as a regex to the full output text. PASS if zero matches found.
   structure_present	Search the output text for the literal string params.marker. PASS if found.
   structure_absent	Search the output text for the literal string params.marker. PASS if NOT found.

2. Record the result:

   • PASS: the assertion result is PASS with the measured value (e.g., word count = 1247)
   • FAIL: the assertion result is FAIL with the measured value and the expected condition

Step 3 — Run llm-judge assertions

For each assertion where type: llm-judge:

1. Construct the evaluation prompt:

   {assertion.prompt}
   
   ---
   OUTPUT TO EVALUATE:
   {full content of output_file}

2. Submit the prompt. Parse the response for a binary verdict: PASS or FAIL.

3. Extract the one-sentence explanation from the response.

4. Record the result:

   • PASS: result is PASS with the LLM's explanation
   • FAIL: result is FAIL with the LLM's explanation

Step 4 — Compile score report

After all assertions are evaluated, compile the score report:

1. Count total assertions run and total assertions passed.
2. List all failed assertions with their IDs, descriptions, and failure details.
3. Produce the structured output (see Outputs section).

---

Quality Checks

• Every assertion in the evals.yaml is evaluated — no assertion is skipped silently
• Each assertion result records its measured value or LLM rationale, not just PASS/FAIL
• The total score is expressed as N/total (e.g., 4/5)
• Failed assertions are listed with enough detail to understand what was measured and why it failed
• The score report is structured such that an agent can parse it programmatically (not free prose)
• If any assertion has an unsupported check value: report as ERROR, do not skip silently

---

Outputs

The skill produces a score report in the following structured Markdown format:

# Eval Report: {skill name} — {evals.yaml version}

**Output file:** {output_file path}
**Evals file:** {evals_file path}
**Run date:** {ISO 8601 date}
**Score:** {N}/{total} assertions passed

---

## Results

| ID | Type | Description | Result | Details |
|----|------|-------------|--------|---------|
| A01 | code | Word count within ±15% of target | PASS | 1247 words (range: 1020–1380) |
| A02 | code | Kill list word 'leverage' absent | FAIL | 2 matches found |
| A03 | llm-judge | Post stands alone without prior context | PASS | "The post opens with a clear hook and requires no prior context to understand." |

---

## Failed Assertions

### A02 — Kill list word 'leverage' absent
- **Type:** code
- **Check:** regex_not_match
- **Pattern:** `\bleverag(e|ing|ed)\b`
- **Result:** FAIL — 2 matches found at positions [line 4, line 11]

The score report may also be emitted as structured YAML if the invoking agent requires machine-readable output:

eval_report:
  skill: content-draft
  evals_version: "1.0"
  output_file: {path}
  evals_file: {path}
  run_date: {ISO 8601}
  score:
    passed: 4
    total: 5
    ratio: "4/5"
  results:
    - id: A01
      type: code
      description: "Word count within ±15% of target"
      result: PASS
      details: "1247 words (range: 1020–1380)"
    - id: A02
      type: code
      description: "Kill list word 'leverage' absent"
      result: FAIL
      details: "2 matches found"
  failed_assertions:
    - id: A02
      description: "Kill list word 'leverage' absent"
      type: code
      check: regex_not_match
      pattern: "\\bleverag(e|ing|ed)\\b"
      details: "2 matches found at positions [line 4, line 11]"

---

Non-Goals

This skill must NOT:

• Modify the output file being evaluated
• Modify the source skill whose output is being evaluated
• Invoke any other skill (skills never chain)
• Make recommendations about what to change in the skill or its output
• Generate an evals.yaml file (that is agent work in the Skill Optimize protocol)
• Compare scores across multiple runs (that is agent orchestration)
• Propose a verdict on whether the skill should be updated (that is a human decision)

No silent skips. Every assertion produces an explicit PASS, FAIL, or ERROR result.